Github user chenghao-intel commented on the pull request:
https://github.com/apache/spark/pull/5154#issuecomment-85569221
Verified the code change by the following micro-benchmark
```scala
import org.apache.spark.sql.catalyst.expressions._
import org.apache.spark.sql.types._
case class Floor(child: Expression) extends UnaryExpression with Predicate {
override def foldable = child.foldable
def nullable = child.nullable
override def toString = s"Floor $child"
override def eval(input: Row): Any = {
child.eval(input) match {
case null => null
case ts: Int => ts - ts % 300
}
}
}
object T {
def benchmark(count: Int, expr: Expression): Unit = {
var i = 0
val row = new GenericRow(Array[Any](123, 21, 42))
val s = System.currentTimeMillis()
while (i < count) {
expr.eval(row)
i += 1
}
val e = System.currentTimeMillis()
println (s"${expr.getClass.getSimpleName} -- ${e - s} ms")
}
def main(args: Array[String]) {
def func(ts: Int) = ts - ts % 300
val udf0 = ScalaUdf(func _, IntegerType, BoundReference(0, IntegerType,
true) :: Nil)
val udf1 = Floor(BoundReference(0, IntegerType, true))
benchmark(1000000, udf0)
benchmark(1000000, udf0)
benchmark(1000000, udf0)
benchmark(1000000, udf1)
benchmark(1000000, udf1)
benchmark(1000000, udf1)
}
}
```
Without the code change it outputs
ScalaUdf -- 1183 ms
ScalaUdf -- 887 ms
ScalaUdf -- 929 ms
Floor -- 49 ms
Floor -- 15 ms
Floor -- 21 ms
With the code change, it outputs
ScalaUdf -- 382 ms
ScalaUdf -- 255 ms
ScalaUdf -- 247 ms
Floor -- 27 ms
Floor -- 6 ms
Floor -- 8 ms
Conclusions:
* The code change will improve the performance of scala udf by 2-3x
* Scala UDF is in very low performance compare to the built-in type of
Expression.
We probably need to provide more efficient way of UDF extension interface.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
